Dual-Direction Co-attention Graph Convolutional Networks for Rumor Detection on Social Media

Abstract

There is a flood of false information on social media. Because of the long tail effect of rumors, certified false information still appears on social media frequently, bringing adverse effects. It is necessary to block the spread of false information in time. However, some existing studies on event-level detection limit the definition of problems to binary classification, and other models are either limited to mining features from information flows arranged in time series or to strict topological relationships, In order to be more practical, we adopt a finer-grained rumor classification standard and propose a dual-direction co-attention graph convolutional network, which makes every tweet in an event capture the long-distance dependence of all tweets on it, and enhances the features fusion from the aspect of propagation and dispersion Extensive experiments on two Twitter datasets and one Weibo dataset show that our method achieves better performance than the most advanced ones. It has a direct impact on practitioners: it can connect with the existing rumor detection mechanism, screen out the information which needs to be checked by domain experts, and achieve cost reduction and efficiency increase accurately and efficiently.